cureAssess

cureAssess helps you decide whether a cure model is appropriate for right-censored survival data — that is, whether the data plausibly contain a fraction of subjects who will never experience the event of interest.

Fitting a cure model when follow-up is too short, or when there is no real cured fraction, produces biased estimates. cureAssess packages the standard screening and diagnostic procedures for that decision into one workflow.

Installation

Install the released version from CRAN:

install.packages("cureAssess")

Or the development version from GitHub:

# install.packages("pak")
pak::pak("GeethanjaleeM/cureAssess")

The two-stage workflow

Stage 1 — screening. Standardize the data, fit matched cure and non-cure parametric models, and rank them by AIC. If the smallest-AIC model is a cure model, that is initial support for cure modeling.

Stage 2 — diagnostics. Run the formal checks for sufficient follow-up and for a cured fraction: the Maller-Zhou statistics, Shen’s test, an immune summary of the Kaplan-Meier tail, and the RECeUS method.

cure.appropriateness() runs both stages in one call.

Example

library(cureAssess)
library(survival)

res <- cure.appropriateness(
  data = gbsg,
  time = "rfstime",
  status = "status",
  time_scale = "days_to_years",
  plot_km = FALSE,
  run_tests = "yes"
)

res
#> 
#> Cure model appropriateness analysis
#> -----------------------------------
#> Best model by AIC: loglogistic_cure 
#> Best model type: cure 
#> 
#> Initial decision:
#> The model with the smallest AIC is a cure model (loglogistic_cure). This provides initial support for cure model appropriateness. 
#> 
#> RECeUS distribution used: llogis 
#> Based on cure model: loglogistic_cure 
#> 
#> Testing status:
#> Tests were run because `run_tests = "yes"`. RECeUS used distribution: llogis. 
#> 
#> Final recommendation:
#> The model with the smallest AIC is a cure model (loglogistic_cure). This provides initial support for cure model appropriateness. Additional cure-appropriateness diagnostics were run for further evaluation.

The AIC comparison behind the decision is available directly:

head(res$screening$aic_table[, c("model", "model_type", "AIC")], 4)
#>              model model_type      AIC
#> 1 loglogistic_cure       cure 1719.696
#> 2       gamma_cure       cure 1723.476
#> 3      loglogistic   non-cure 1731.346
#> 4     weibull_cure       cure 1733.922

Individual diagnostics can also be run on their own, on data prepared by prepare.surv.data():

dat <- prepare.surv.data(
  data = gbsg,
  time = "rfstime",
  status = "status",
  time_scale = "days_to_years"
)

mz.test(dat)
#> 
#> Maller-Zhou test statistic (1994)
#> ---------------------------------
#> Statistic: 0.0494606 
#> Alpha: 0.05 
#> Interpretation: Since the Maller-Zhou statistic (0.0495) is less than alpha = 0.05, there is evidence of sufficient follow-up, supporting cure model appropriateness.
receus.method(dat, dist = "lnorm")
#> 
#> RECeUS Cure Model Assessment
#> ----------------------------
#> Distribution: lnorm 
#> Tau: 7.2799 
#> 
#> Estimated cure fraction (pi_hat): 0.2775 
#> Remaining uncured ratio (r_hat): 0.4083 
#> 
#> Decision: Follow-up insufficient for cure modeling 
#> 
#> Interpretation:
#> The remaining uncured ratio (r_hat = 0.4083 ) is greater than or equal to 0.05. This suggests that a large proportion of uncured subjects remain censored at the end of follow-up, indicating insufficient follow-up to reliably estimate a cure fraction.

Note what the two stages say here. AIC preferred a cure model for these data, but RECeUS reports that too large a share of uncured subjects is still censored at the end of follow-up. A better AIC fit is not on its own evidence that a cure fraction can be estimated reliably — which is why the screening stage alone is not enough.

For a worked comparison of a data set where a cure model is supported (nwtco) against one where it is not (gbsg), see the introductory vignette:

vignette("introduction", package = "cureAssess")

Interpreting the output

These diagnostics are descriptive aids, not a single decision rule. Sufficient follow-up tests ask whether the study ran long enough to tell a genuine cure fraction apart from a plateau caused by censoring. Read them together, and alongside subject-matter knowledge about whether cure is clinically plausible.

References

License

MIT © Geethanjalee Mudunkotuwa